Autonomous Platoon Control With Integrated Deep Reinforcement Learning and Dynamic Programming

Autonomous vehicles in a platoon determine the control inputs based on system state information collected and shared by Internet of Things (IoT) devices. Deep reinforcement learning (DRL) is regarded as potential method for car-following has been mostly studied to support single following vehicle. However, it more challenging learn an efficient policy with convergence stability when there are multiple platoon, especially unpredictable leading vehicle behavior. In this context, we adopt integrated DRL dynamic programming (DP) approach autonomous policies, which embeds deep deterministic gradient (DDPG) algorithm into finite-horizon value iteration framework. Although DP framework can improve performance DDPG, limitations lower sampling training efficiency. article, propose algorithm, namely, finite-horizon-DDPG sweeping through reduced space using stationary approximation (FH-DDPG-SS), uses three key ideas overcome above limitations, i.e., transferring network weights backward time, earlier time steps, space. order verify effectiveness FH-DDPG-SS, simulation real driving data performed, where FH-DDPG-SS compared those benchmark algorithms. Finally, safety string demonstrated.